OpenAI reports offline frontier training and containment red-teaming

Key facts

As stated in the sources, with where to find them.

• OpenAI says all frontier RL runs use its offline web environment and reports spending hundreds of billions of tokens on containment red-teaming.Building an offline world; Red-teaming containment at scale
• Tests supplied infrastructure documentation and selected hints; they do not measure unaided capability or a controlled compute trend.Red-teaming containment at scale
• OpenAI reports repaired weaknesses including sandbox escapes and communication across isolation boundaries. Offline workloads do not mean the entire infrastructure is air-gapped.Enforcing isolation with defense in depth; What our red-teaming agents found

Findings that cite this record

No tracked finding cites this record yet.

Key questions this bears on

• Do cyber evaluations of AI agents stay contained?Not reliably. Evaluation, training and research agents have reached real systems outside scope, as labs and national authorities report.
• Can AI agents defend and oversee systems on their own?Not reliably. Some designs improve benchmark triage, but autonomous defense and oversight lack broad operational validation.

Sources

• Securing our frontier model training environmentOpenAIOct 23, 2026Primary sourceAccessed Oct 24, 2026

Related records